Research map / 3D vision / Depth and stereo
UltraLED: Learning to See Everything in Ultra-High Dynamic Range Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UltraLED, a two-stage single short-exposure RAW framework (ratio-map exposure correction followed by brightness-aware RAW denoising) for ultra-high dynamic range scenes, with a new 9-stop synthesized training pipeline and a static UHDR dataset.
Problem
Reconstructing ultra-high dynamic range scenes with bimodal exposure distributions (bright highlights and very dark shadows) without ghosting from multi-frame bracketing, while recovering detail and suppressing noise in dark regions.
Approach
Decouple UHDR reconstruction from a single short-exposure RAW frame into a learned per-pixel ratio map for exposure correction and a RAW denoiser conditioned on Gaussian-encoded ratio values, plus a brightness-aware noise term in the training synthesis.
Key results
- On the UHDR dataset at x50 ratio, UltraLED reaches 27.591 PSNR / 0.898 SSIM, versus 25.942 PSNR for the best RAW baseline (ELD).
- Ablation shows the RAW domain outperforms RGB, with the decoupled RAW pipeline at x50 reaching 27.598 PSNR compared with 22.937 for the RGB decoupled variant.
- A user study with 122 participants found users consistently preferred UltraLED over baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the two-stage RAW pipeline rather than one-stage RAW reconstruction (ablation) | UltraLED · UHDR dataset ×50, ×100, and ×200 | PSNR: 25.137 -> 27.598; 24.896 -> 27.327; 24.632 -> 27.091 | Helped |
| Use RAW rather than RGB input (ablation) | UltraLED · UHDR dataset ×50 | PSNR: 22.937 -> 27.598 | Helped |
| Add brightness-aware noise modeling (ablation) | UltraLED · UHDR dataset ×50 | PSNR: 26.301 -> 27.598 | Helped |
| Add ratio-map encoding (ablation) | UltraLED · UHDR dataset ×50 | PSNR: 27.062 -> 27.598 | Helped |
| Use the proposed Gaussian ratio-map encoding instead of no encoding (ablation) | UltraLED · UHDR dataset ×50, ×100, and ×200 | PSNR: 27.062 -> 27.598; 26.893 -> 27.327; 26.734 -> 27.091 | Helped |
| Use the proposed encoding instead of one-hot encoding (ablation) | UltraLED · UHDR dataset ×50, ×100, and ×200 | PSNR: 27.463 -> 27.598; 27.201 -> 27.327; 26.982 -> 27.091 | Helped |
| Use the proposed encoding instead of position encoding (ablation) | UltraLED · UHDR dataset ×50, ×100, and ×200 | PSNR: 26.893 -> 27.598; 27.160 -> 27.327; 26.781 -> 27.091 | Mixed |
| Use Restormer instead of U-Net as denoising backbone (ablation) | UltraLED · UHDR dataset ×50, ×100, and ×200 | PSNR: 27.598 -> 27.982; 27.327 -> 27.732; 27.091 -> 27.437 | Helped |
| Add ratio-map encoding to PG (ablation) | PG · ELD dataset ×1, ×10, ×100, and ×200 | PSNR: 54.974 -> 56.032; 51.228 -> 51.593; 42.029 -> 42.132; 38.059 -> 38.413 | Helped |
| Add ratio-map encoding to ELD (ablation) | ELD · ELD dataset ×1, ×10, ×100, and ×200 | PSNR: 53.092 -> 54.141; 50.940 -> 51.375; 45.529 -> 45.646; 43.010 -> 43.079 | Helped |
| Add ratio-map encoding to SFRN (ablation) | SFRN · ELD dataset ×1, ×10, ×100, and ×200 | PSNR: 53.478 -> 54.662; 51.231 -> 51.805; 45.631 -> 45.841; 43.013 -> 43.021 | Helped |
| Use UltraLED instead of HDRTVNet++ | UHDR reconstruction · UHDR dataset | PSNR: 25.344 -> 27.591 | Helped |
Surprising findings
- Decoupling exposure correction and denoising helps only in RAW (RGB decoupling slightly lowers PSNR at x100 and x200).
Limitations
- Strong dependence of the noise model on camera-specific parameters; cross-camera performance drops, so retraining per camera type is needed.
- Evaluation dataset is mostly indoor controlled scenes (24 scenes), with outdoor scenes only qualitative or user-study based.
- Requires RAW input, so it does not apply to RGB-only capture.
Experimental setup
Datasets: RawHDR SID ELD
Benchmarks: UHDR dataset SID ELD
Metrics: PSNR SSIM LPIPS
Baselines: PG ELD LED RetinexMamba EnlightenGAN Zero-DCE Kind HDRUNet HDRTVNet HDRTVNet++
Builds on: SID ELD LED RAWNeRF Exposure Fusion
Closest papers by meaning
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.89)
- NEC-Diff: Noise-Robust Event-RAW Complementary Diffusion for Seeing Motion in Extreme Darkness (CVPR 2026 · similarity 0.85)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.85)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.85)
- Exposure Bracketing Is All You Need For A High-Quality Image (ICLR 2025 · similarity 0.84)
- Unveiling Details in the Dark: Simultaneous Brightening and Zooming for Low-Light Image Enhancement (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.